On Sat, 8 Aug 2015, Christian Kujau wrote:

[Adding linux-ide@vger.kernel.org]

On Fri, 7 Aug 2015, Christian Kujau wrote:

quoted

this PowerBook G4 was running 3.16 for a while but now I wanted to upgrade 
to latest mainline. However, during bootup the following happens:

===============================
[    2.237102] ata1: PATA max UDMA/100 irq 39
[    2.401708] ata1.00: ATA-8: SAMSUNG HM061GC, LR100-10, max UDMA/100
[    2.401764] ata1.00: 117231408 sectors, multi 16: LBA48 
[    2.417633] ata1.00: configured for UDMA/100
[   44.918102] ata1.00: exception Emask 0x0 SAct 0x0 SErr 0x0 action 0x0
[   44.920452] ata1.00: failed command: READ DMA
[   44.922725] ata1.00: cmd c8/00:88:64:c2:12/00:00:00:00:00/e0 tag 0 dma 69632 in
[   44.927257] ata1.00: status: { DRDY }
[   49.971784] ata1.00: qc timeout (cmd 0xec)
[   49.976529] ata1.00: failed to IDENTIFY (I/O error, err_mask=0x4)
[   49.978908] ata1.00: revalidation failed (errno=-5)
[   55.019662] ata1: link is slow to respond, please be patient (ready=0)
[   60.007677] ata1: device not ready (errno=-16), forcing hardreset
[   60.012670] ata1: soft resetting link
[   60.193638] ata1.00: configured for UDMA/100
[   60.196158] ata1.00: device reported invalid CHS sector 0
[   60.198610] ata1: EH complete
===============================

This happens only once, but systemd thinks there's a hard problem and will 
drop to a recovery shell. I can start sshd and login remotely and then the 
system appears to be running just fine.

I played around with libata* kernel parameters, the only "success" I had 
was with libata.dma=0 - which disables DMA and the system booted without 
the error. But of course the disk throughput was much slower - is there a 
way to enable DMA again once the system is booted? "hdparm -d" would 
return HDIO_SET_DMA, of course[0].

Tried something more drastic and disabled libata completely and enabled 
CONFIG_IDE (and CONFIG_BLK_DEV_IDE_PMAC) again and a similar error appears 
(sometimes) during bootup:

[   39.971392] ide-pmac lost interrupt, dma status: 8480
[   39.972704] hda: lost interrupt
[   39.973951] hda: dma_intr: status=0xd8 { Busy }
[   39.975231] hda: possibly failed opcode: 0x25
[   39.978855] hda: DMA disabled
[   40.019388] ide0: reset: success

But the host seems to recover more quickly and systemd wasn't thrown off 
by the small ATA delay. But DMA got disabled again :-\

Ideas welcome! :-)

Christian.

[0] https://ata.wiki.kernel.org/index.php/Libata_FAQ

quoted

This happened in 4.2.0-rc5 so I went back a few versions and found that
4.1-rc5 was OK (the error does not show up and the system boots just fine)
and 4.1-rc6 is not.

After more digging around I noticed that the same error (with 
changed wording) happens with a Debian 3.16.0-4-powerpc kernel - so it
doesn't appear to be a recent regression as I suspected at first:

==================================
[   46.907147] ata1: drained 572 bytes to clear DRQ
[   46.907166] ata1.00: exception Emask 0x0 SAct 0x0 SErr 0x0 action 0x0
[   46.908419] ata1.00: failed command: READ DMA
[   46.909058] ata1.00: cmd c8/00:80:9c:f9:60/00:00:00:00:00/e0 tag 0 dma 65536 in
         res 40/00:fe:00:00:00/00:00:00:00:00/40 Emask 0x20 (host bus error)
[   46.910303] ata1.00: status: { DRDY }
[   46.970579] ata1.00: configured for UDMA/100
[   46.971853] ata1.00: device reported invalid CHS sector 0
[   46.972524] ata1: EH complete
==================================

Also, the error cannot repduced as reliably as I thought: sometimes, the 
machine just boots w/o a hitch - and that might be the reasons why my 
bisect attempts failed and incorrectly blamed totally unrelated commits: 
after each "git bisect {good,bad}" (+compiling) I rebooted but there was a 
chance that the system came up just fine / showed the same ATA error and 
thus falsified the git-bisect results.

I noticed that with this Debian 3.16 kernel, it happens less often when I 
use the "irqpoll" option. But with 4.2-rc5 this doesn't seem to help much, 
the system still hangs during boot but continues after the "EH complete " 
message. And it doesn't appear afterwards, I can read from my root disk 
just fine and a long SMART check also comes back fine.

Because the error only appears to happen on the very first access after a 
reboot, I tried to boot with rootdelay=30 - but of course then it just 
waits "before" accessing the root disk. I'd need a magic option to wait a 
few seconds "after" the first disk access, so that the boot framework 
("systemd") won't be thrown off when /dev/sda isn't responding as fast as 
expected.

What _does_ seem to help a bit was to disable the the swap device, which 
is configured as an encrypted dm-device here - and systemd was almost 
always stumbling over this particular service during bootup. Because of 
the ATA timeout, the dm-device could not be setup correctly and systemd 
would bail out and drop me into a recovery shell. Without the swap device, 
systemd would skip setting up swap and boot just fine (most of the time) 
and I can setup swap once the system has been booted. So...there's that.

It's still a mystery to me why /dev/sda is only behaving weird on its 
first access.

I've cc'ed linux-ide, maybe somebody has an idea on that?

dmesg & .config:  http://nerdbynature.de/bits/v4.1-rc6/

Thanks,
Christian.

quoted

Unfortunately a git-bisect between these two versions went completly off 
the charts, I don't know what happened here:
==================================
first bad commit:

0fa372b6c95013af1334b3d5c9b5f03a70ecedab is the first bad commit
commit 0fa372b6c95013af1334b3d5c9b5f03a70ecedab
Author: Takashi Iwai [off-list ref]
Date:   Wed May 27 16:17:19 2015 +0200

    ALSA: hda - Fix noise on AMD radeon 290x controller
==================================

I don't have this driver (or ALSA) even selected. I can reproduce this 
error pretty reliably and I'd like to attempt another git-bisect
run when I'm more awake. But maybe somebody recognizes this error and
has a hint where this could come from?

dmesg & .config:  http://nerdbynature.de/bits/v4.1-rc6/

Thanks,
Christian.
-- 
BOFH excuse #225:

It's those computer people in X {city of world}.  They keep stuffing things up.

-- 
BOFH excuse #263:

It's stuck in the Web.

-- 
BOFH excuse #316:

Elves on strike. (Why do they call EMAG Elf Magic)